The solution of the wave equation by wavelets basis approximation Pascal Joly

∗

Dimitri Komatitsch

†

Jean-Pierre Vilotte

‡

Enumath conference, September 18-22, 1995, Paris, France Abstract This paper is the first development of a work concerning the use of wavelets basis in the numerical solution of the one-dimensional wave equation, that we consider as a good introduction to more realistic two or three-dimensional seismological problems, as it is important to check the accuracy of the tools required for intensive numerical simulation. The first section of this paper is devoted to the theoretical background : the multiresolution analysis ; in the second section we introduce some associated numerical algorithms and in the last section, we present some numerical results. AMS(MOS) subject classifications. 65F10

1

Multiresolution Analysis

1.1

Introduction

The multiresolution analysis frame for variational spaces such as L2 (R), has been recently developed by I. Daubechies [6], S. Mallat [14] [15] and Y. Meyer [17] [18]. A multiresolution analysis of L2 (R) is, by definition, an increasing sequence {Vj }j ∈Z of closed subspaces having the following properties: (1) (2) (3) (4) (5)

Vj ⊂ Vj+1 f ∈ Vj ⇔ f 1 ∈ Vj+1 T Vj = {0} Sj∈Z V is dense in L2 (R) j∈Z j there exists a function g in V0 , such that {gk }k∈Z is a Riesz basis of V0

∗

C.N.R.S. Laboratoire d’Analyse Num´erique. Universit´e Pierre et Marie Curie. Paris Laboratoire de Sismologie. Institut de Physique du Globe. Paris ‡ Laboratoire de Sismologie. Institut de Physique du Globe. Paris †

1

2

Wavelets for the wave equation where by definition ∀f ∈ L2 (R), ∀x ∈ Rf k (x) = 2k/2 f (2k x) ∀g ∈ L2 (R), ∀x ∈ R gk (x) = g(x − k)

1.2

The scaling function

Let us denote ϕ the function defined by ϕ(ω) ˆ =(

X

|ˆ g (ω + 2lπ)|2 )−1/2 gˆ(ω)

l∈Z

this function satisfies (see [5]) X

|ϕ(ω ˆ + 2lπ)|2 = 1

l∈Z

hence ∀k ∈ Z

Z

ϕ(x) × ϕ(x − k) dx = δ0,k

and then {ϕk }k∈Z is an orthonormal basis of the subspace V0 . ϕ is the so-called scaling function, and it is used to define the subspaces Vj : first define the functions ϕjk by ϕjk (x) = 2j/2 ϕ(2j x − k)

∀x ∈ R, ∀j, k ∈ Z

For some fixed j ∈ Z, each orthonormal family {ϕjk }k∈Z generates a subspace Vj , and the sequence {Vj }j ∈Z is a multiresolution analysis of L2 (R).

1.3

The wavelet function

Now for all j ∈ Z, let Wj be the orthogonal complementary of Vj in Vj+1 : Vj+1 = Vj ⊕ Wj , then again there exists a function ψ such that {ψk }k∈Z is an orthonormal basis of W0 . Furthermore, let us define the functions ψkj by ψkj (x) = 2j/2 ψ(2j x − k)

∀x ∈ R, ∀j, k ∈ Z

each family {ψkj }k∈Z is an orthonormal basis of Wj , and ψ is the so-called wavelet function.

1.4

The bounded domain case

The multiresolution analysis is available in the case of a bounded domain (see for instance [19] and [2]). We are interested in the periodic case on an interval. In [17], Y. Meyer has introduced the periodic multiresolution analysis: ∀ x ∈ [0, 1] j ≥ 0,

0 ≤ k < 2j

∀ x ∈ [0, 1] j ≥ 0,

0 ≤ k < 2j

ϕjk (x) = 2j/2

X

ϕ(2j (x + z − k)) z∈ Z X j j/2 ψ(2j (x + z − k)) ψk (x) = 2 z∈Z

. .

3

Wavelets for the wave equation

The functions ϕjk and ψkj generate subspaces Vj et Wj and define a multiresolution analysis of L2 (T ), where T is the torus R/Z, and L2 (T ) is a set of 1-periodical functions. We shall use this particular approach in the following; see also [20] for a practical use of these wavelets. Note that the dimension of the subspace Vj is 2j , so the multiresolution analysis defines a Galerkin approximation of L2 (T ). For numerical experiments, we consider a fixed integer p, the dimension of the corresponding subspace Vp is 2p , the associated multiresolution is Vp = V0 ⊕ W0 ⊕ W1 . . . ⊕ Wp−1

1.5

Some examples of wavelet functions

It is not the aim of this paper to list all available wavelet functions, we just recall some famous ones : the Haar’s wavelet function [9] is certainly the oldest example ; there exist also the fast decreasing Meyer’s wavelet [17], the compact supported Daubechies’ wavelets [6] and the Battle-Lemari´e’s wavelets with exponential decrease [3], [11]. In the case of G. Battle and P.G. Lemari´e’s wavelets, both functions ϕ and ψ are defined by their Fourier transform (see also [7]) sinm (πω) 1 ϕ(ω) ˆ = (πω)m [Pm−1 (sin2 (πω))]1/2 ˆ ψ(ω) =



e−iωπ Pm−1 (cos2 (πω/2)) (πω/2)m Pm−1 (sin2 (πω/2))Pm−1 (sin2 (πω))

1/2

sin2m (πω/2)

with Pm−1 (sin2 (z)) = (sin(z))2m

1 ( )2m (z + l π) l∈Z X

Figure 1 presents the function ψ00 , which is used to define all the wavelets functions (level p = 9). It is the only basis function of subspace W0 . Figures 2 and 3 show the two basis functions of W1 : ψ01 and ψ11 (note that the real magnitude of the signal is modified). It is easy to see from figures 4 to 7, that the support length of the basis functions of Wj decreases as j increases. One should also note that the 2j basis functions of Wj are obtained from ψ0j by x-translations of length k × 2−j (Figures 7 and 8).

4

Wavelets for the wave equation

Figure 1 : the only wavelet of W0

Figure 5 : one wavelet of W3

Figure 2 : the first wavelet of W1

Figure 6 : one wavelet of W5

Figure 3 : the second wavelet of W1

Figure 7 : one wavelet of W6

Figure 4 : one wavelet of W2

Figure 8 : another wavelet of W6

5

Wavelets for the wave equation

2

Algorithms

2.1

Introduction

We present in this section some algorithms that are useful for the different steps of the calculations. The most important step is the determination of the wavelet coefficients, that is the components of some given function f ∈ L2 in the subspaces Wj , with 0 ≤ j < p, for a multiresolution with p levels. Although the following procedure is available in the general case, we give the formulas in the case of periodic boundary conditions on the interval [0, 1]. The very first step is a to get a regular sampling of the function f , that is the 2p values fk = f (k/2p ) (0 ≤ k < 2p ); then an interpolation step has to be performed, i.e. find one element f˜p ∈ Vp ⊂ L2 such that f˜p (k/2p ) = f (k/2p )

∀k, 0 ≤ k < 2p

2.2

Interpolation p

We look for the 2 coefficients

cpk

p −1 2X

such that f˜p =

cpk ϕpk .

k=0

In the case of the Battle-Lemari´e’s wavelet, there exists a function S ∈ Vp satisfying ∀k, 0 ≤ k < 2p

p ˜ S(k/2 ) = δ0,k

so that the coefficients are defined by (see [20]) cpk =

∀k, 0 ≤ k < 2p

p −1 2X

fl Lp (k − l)

l=0

where ∀k, 0 ≤ k < 2p

2.3

Lp (k) =

Z

0

1

S(x)ϕpk (x) dx

Decomposition

Now we get the 2p coefficients cpk defining an element fp ∈ Vp ; the next step is to obtain the similar coefficients cjk in each subspace Vj (0 ≤ j < p). This is the decomposition step, for which we use the Mallat transform (see [14]). From the relation Vj = Vj−1 ⊕ Wj−1 we deduce for any j (0 < j < p) f˜ ∈ Vj−1 f˜j = f˜j−1 + g˜j−1 with { j−1 g˜j−1 ∈ Wj−1

6

Wavelets for the wave equation The 2j coefficients ckj−1 and dkj−1 satisfy f˜j−1 =

2j−1 X−1

ckj−1

ϕkj−1

and

g˜j−1 =

k=0

2j−1 X−1

dkj−1 ψkj−1

k=0

the dkj−1 are the so-called wavelet coefficients of level j − 1. If we define two discrete 2j -periodic filters (Gj (n))n∈Z and (Hj (n))n∈Z by 0≤j